Communications of the ACM
Scan Your Doodles to Find the Perfect Matching Photo Online
New software developed at the Georgia Institute of Technology searches for a photographic match to a hand-drawn sketch.
Credit: Vincent Migeat/Agence VU/Camera Press
Georgia Institute of Technology (Georgia Tech) researchers say they have developed software that can scan hand-drawn sketches and search for a photograph that looks just like them.
Georgia Tech professor James Hays says the program is an exciting step toward a search engine based on drawings.
The researchers recruited 664 workers on Amazon Mechanical Turk to draw sketches. A photo was randomly selected from a stack of thousands and then shown to a worker for two seconds; each picture fell into one of 125 categories of recognizable objects. The worker then drew what they had seen from memory, with the whole group spending nearly 4,000 hours sketching.
The program matched the sketches to the original photographs using two neural networks. One of the networks analyzed the sketches, while the other evaluated the photos. The two algorithms then determined which pairs were most similar.
During testing, the program correctly matched the sketch to the photograph 37 percent of the time. Meanwhile, humans completed the same task correctly about 54 percent of the time, which is not an insurmountable goal for artificial intelligence systems, according to Hays.
The researchers will present their work in July at the ACM SIGGRAPH 2016 conference in Anaheim, CA.
From New Scientist
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found